Browsing by Author "Dr. Wesley E. Snyder, Member"

Filter results by typing the first few letters
Now showing 1 - 2 of 2
  • No Thumbnail Available
    A Novel Focused Local Learning Wavelet Network with Application to In Situ Monitoring During Selective Silicon Epitaxy
    (2001-04-10) Rying, Eric Arne; Dr. Griff L. Bilbro, Chair; Dr. F. Yates Sorrell, Co-Chair; Dr. Mehmet C. Ozturk, Member; Dr. J.C. Lu, Member; Dr. Winser E. Alexander, Member; Dr. Wesley E. Snyder, Member
    This dissertation reports results which pioneered the novel application of wavelets to processes and issues critical to semiconductor manufacturing. This work is especially germane given: 1) the rising costs and complexity of manufacturing; 2) the increasing deluge of data provided by process andequipment sensors that are developed, in part, to address these costs; and 3) the lack of adequate tools for handling the information overload. The sheer quantity of data is currently outstripping conventional means of storage and analysis; as a result, an increased fraction of this data must be discarded or incompletely processed. Ultimately, by developing advanced tools and methodologies to address these issues, cost and performance-driven decisions can be made in a timely and cost-effective manner. The content of this dissertation impacts these critical issues through the following specific contributions. Wavelet-based methodologies were developed for the localized modeling and compression of key process-relevant information, especially information pertaining to the detection of equipment and process faults. These methodologies significantly contribute to the following general problems: 1) identification of specific local features in potentially large nonstationary data sets; 2) compression of entire data sets consisting of disjoint smooth and nonsmooth segments; and 3) improvement in the quality of modeling for important local features carrying key information. This thesis addressed these problems using wavelet networks, as the tool, and the following novel approaches: 1) a novel objective function that incorporates both global and local error as well as model parsimony; 2) a novel adaptive network initialization scheme; 3) a novel application of the wavelet transform modulus maxima (WTMM) representation to help determine and prioritize the set of local features; 4) a new network construction procedure; and 5) a new errorspace analysis (ESA) technique, to assist in the visualization of both local and global network approximation errors during wavelet network construction. Information gathered in situ using a quadrupole mass spectrometer (QMS) sensor facilitated the development of a novel and unique approach for monitoring the selective film thickness during selective growth of silicon epitaxy. Moreover, this approach is applicable to the detection of one particularly critical process fault --- loss of selectivity. In particular, since QMS sensors are currently in widespread use in most vacuum systems, this technique represents a viable, cost-effective solution to selectivity loss detection that appears readily transferable to other process chemistries. In general, by correlating similar in situ process metrics (e.g., signal area) to ex situ process observables (e.g., thin-film thickness), contributions are made to: 1) thin-film metrology and metal oxide semiconductor field effect transistor (MOSFET) gate-stack engineering; 2) run-to-run detection of aberrant signalmodes, including process and equipment faults; and 3) decision-making using wavelet-compressed spaces. The methodologies developed in this thesis appear to be broadly applicable to otherdisciplines and fields of endeavor, including intelligent manufacturing and information technology.
  • No Thumbnail Available
    Pattern Analysis, Tracking and Control for Autonomous Mobile Robots Using Neural Networks
    (1998-10-09) Janet, Jason Andre; Dr. Mark W. White, Chair; Dr. Michael G. Kay, Member; Dr. James J. Brickley, Jr., Member; Dr. John C. Sutton III, Member; Dr. Wesley E. Snyder, Member; Dr. William D. Allen, Member
    Autonomous vehicles require that all on-board processes be efficient in time, complexity and data storage. Infact, an ideal system employs multi-funcitonal models where ever possible. The research documented hereproposes that the Region-Feature Neural Network (RFNN) and the Hyper-Ellipsoid Clustering (HEC)Kohonen neural network (or HECNN) are viable pattern analysis and control engines that contribute to thesolution of a variety of problems. The theoretical development of the RFNN and HECNN, along with several proof-of-concept applications are presented in detail. The RFNN is a feed-forward, back-propagation model that is more general than standard textbook models because it also accomodates receptive fields and weightsharing. The RFNN uses a modified version of adaptive learning rates, called "shocking" to reduce training time and maintain stability. Small-scale benchmark problems like the XOR and XOP problems are used to demonstrate the utility of the "shocking" model. Due to its modularity, the RFNN allows the user to construct flexible, multi-layered, feed-forward architectures as well as add to and prune from an architecture even aftertraining has begun. The RFNN also permits the user to include previously learned features, called "analogies" to further expedite the training process on new problems or whenever new classes are added. The HECNN is aself-organizing neural network that incorporates hyperellipsoid clustering by use of the Mahalanobis distance tolearn elongated shapes and obtain a stochastic measurement of data-node association. The number of nodes canalso be regulated in a self-organizing manner by measuring how well each node models the statistical properties of its associated data. This measurement, called "compactness", determines where and whether to add neuralunits or prune them completely. We make several enhancements to the Kolmogorov-Smirnov compactness test to control the triggering of mitosis and/or pruning. Because fewer nodes are needed for an HECNN than for aKohonen that uses only Euclidean distance, the data size is smaller for the HEC Kohonen, even forhigh-dimensional problems. The large-scale pattern analysis problems presented here for the RFNN includesonar pattern recognition and outdoor landmark recognition. For the HECNN, we focus on sonar pattern recognition and (topographical) map building. Both the RFNN and the HECNN can be generalized to solve orcontribute to the solution of other pattern recognition problems. Both are also multifunctional in that they accommodate compact geometric motion planning (MP), self-referencing (SR) and tracking algorithms.Additionally, we propose the "traversability vector" (t-vector) as an efficient bridge between the HECNN andboth motion planning and self-referencing for mobile robots. As with the RFNN and HECNN, the t-vector is amodular and multi-functional tool that minimizes the computation requirements and data size as it detects path obstructions, Euclidean optimal via points, and geometric beacons, as well as identify which geometric featuresare visible to sensors in environments that can be static or dynamic. Tracking is made possible with Julier andUhlmann's unscented filter. The unscented filter particularly compliments the HECNN in that it performslow-level (non-linear) tracking more efficiently and more accurately than its predecessor, the extended Kalmanfilter (EKF). By estimating and propagating error covariances through system transformations, the unscentedfilter eliminates the need to derive Jacobian matrices. The inclusion of stochastic information inherent to the HECmap rendered the JUKF an excellent tool for our HEC-based map building, global self-localization, motionplanning and low-level tracking.